Large-Scale Diet Tracking Data Reveal Disparate Impacts of Food Environment

Tim Althoff,1∗ Hamed Nilforoshan2, Jenna Hua3, Jure Leskovec2,4 1Allen School of Computer Science & Engineering, University of Washington 2Department of Computer Science, Stanford University 3Stanford Prevention Research Center, Department of Medicine, Stanford University School of Medicine 4Chan Zuckerberg Biohub, San Francisco, CA ∗To whom correspondence should be addressed; E-mail: [email protected]

An unhealthy diet is a key risk factor for chronic diseases including obesity, diabetes, and heart disease1, 2. Limited access to healthy food options may contribute to unhealthy diets3, 4. Studying diets is challenging, typically restricted to small sample sizes, single locations, and non-uniform design across studies, and has led to mixed results on the impact of the food environment5–21. Here we leverage smartphones to track diet health and weight status in a country-wide observational study of 1,164,926 U.S. participants and 2.3 billion food entries to study the independent contributions of fast food and grocery access, income and education on these outcomes. This study constitutes the largest nationwide study examining the impact of the food environment on diet to date, with 300 times more participants and 4 times more person years of tracking than the Framingham Heart Study. We find that higher access to grocery stores, lower access to fast food, higher income and education are independently associated with higher consumption of fresh fruits and vegetables, lower consumption of fast food and soda, and lower likelihood of being overweight/obese, but that these associations vary significantly across predominantly Black, Hispanic, and White zip codes. For instance, within predominantly Black zip codes we find that high income is associated with a decrease in healthful food consumption patterns across fresh fruits and vegetables (6.5%) and fast food (5.5%). Further, high grocery access has a significantly larger association with increased fruit and vegetable consumption in predominantly Hispanic (7.4% increase) and Black (10.2% increase) zip codes in contrast to predominantly White zip codes (1.7% increase). Policy targeted at improving food access, income and education may increase healthy eating, but interventions may need to be targeted to specific subpopulations for optimal effectiveness. 1 1 Introduction Unhealthy diet is the leading risk factor for disability and mortality globally1. Emerging evidence suggests that the built and food environment, behavioral, and socioeconomic cues and triggers significantly affect diet5. Prior studies of the food environment and diet have led to mixed results5–21, and very few used nationally representative samples. These mixed results are potentially due to methodological limitations of small sample size, differences in geographic contexts, study population, and non-uniform measurements of both the food environment and diet across studies. Therefore, research with larger sample size and using improved and consistent methods and measurements is needed7, 22, 23. With ever increasing smartphone ownership in the U.S.24 and the availability of immense geospatial data, there are now unprecedented opportunities to combine various data on individual diets, population characteristics (gender, race and ethnicity), socioeconomic status (income and education), as well as food environment at large scale. Interrogation of these rich data resources to examine geographical and other forms of heterogeneity in the effect of food environments on health could lead to the development and implementation of cost-effective interventions25. Here, we leverage large-scale smartphone-based food journals and combine several Internet data sources to quantify the independent impact of food (grocery and fast food) access, income and education on food consumption and weight status of 1,164,926 subjects across 9,822 U.S. zip codes. This study constitutes the largest nationwide study examining the impact of the food environment on diet to date, with 300 times more participants and 4 times more person years of tracking than the Framingham Heart Study26.

2 2 Methods 2.1 Study Design and Population We conducted a United States countrywide cross-sectional study of participants’ self-reported food intake and body-mass index (BMI) in relation to demographic (education, ethnicity), socioeconomic (income), and food environment factors (grocery store and fast food access) captured on zip code level. Overall, this cross-sectional matching-based study analyzed 2.3 billion food intake logs from U.S. smartphone participants over seven years across 9,822 zip codes (U.S. has total of 41,685 zip codes). Participants were participants of the MyFitnessPal app, a free application for tracking caloric intake. We analyzed anonymized, retrospective data collected during a 7-year observation period between 2010 and 2016 that were aggregated to the zip code level. Comparing our study population to nationally representative survey data, we found that our study population had significant overlap with the U.S. national population in terms of population demographics, education and weight status (Body Mass Index; BMI), but that it was skewed towards women and higher income (Supplementary Table e1). Our matching-based statistical methodology controls for observed biases between comparison groups in terms of income, education, grocery access, and fast food access (Section 2.4). Data handling and analysis was conducted in accordance with the guidelines of the Stanford University Institutional Review Board. 2.2 Study Data We compute outcome measures of food consumption and weight status from 2.3 billion food intake logs by 1,164,926 U.S. participants of the MyFitnessPal (MFP) smartphone application to quantify food consumption across 9,822 zip codes. During the observation period from January 1, 2010 to November 15, 2016, the average participant logged 9.30 entries into their digital food journal per day. The average participant used the app for 197 days. All participants in this sample used the app for at least 10 days. We classified the 2.3 billion food intake entries into three categories of public health interest, fresh fruits and vegetables (F&V), fast food, and sugary non-diet soda, and excluded them from analysis if they did not match these categories. For more details see Supplementary Information. We intentionally use a cross-sectional rather than longitudinal study design, since fine-grained and large-scale temporal data on changes in the food environment were not available. We obtained data on demographic and socioeconomic factors from CensusReporter27. Specif- ically, for each zip code in our data set we obtained median family income, fraction of population

3 with college education (Bachelor’s degree or higher), and fraction of population that is White (not including Hispanic), Black, or Hispanic from the 2010-14 American Community Survey’s census tract estimates27. While data were available only on zip code level, previous studies have shown that area-level income measures are meaningful for health outcomes and describe unique socioeconomic inequities.28 Grocery store access was defined as the fraction of population that is more than 0.5 miles away from a grocery store following the food desert status definitions from the USDA Food Ac- cess Research Atlas29. Contrary to the USDA definition, we found evidence that even in rural zipcodes, the fraction of population greater than 0.5 miles away from grocery stores has the strongest association with food consumption (compared to 10 and 20 miles away), and thus we used 0.5 miles as the threshold in both rural and urban zipcodes (Supplemental Methods: Details on food environment measures). We measured fast food access through the fraction of restaurants that are fast food restaurants within a sample from Yelp, querying the nearest 1000 businesses from the zip code’s center, up to a maximum radius of 40 km (25 miles). See Supplementary Information for details and validation of these objective food environment measures. We will release all data aggregated at a zipcode level in order to enable validation, follow-up research, and use by policy makers. 2.3 Reproducing state-of-the-art measures using population-scale digital food logs To investi- gate the applicability of population-scale digital food logs to study the impact of food environment, income and education on food consumption, we measured the correlation between our smartphone app-based measures and state-of-the-art measures of food consumption including the Behavioral Risk Factor Surveillance System (BRFSS), based on representative surveys of over 350,000 adults in the United States 30, 31, and the Nielsen Homescan data 32, which is a nationally representative panel survey of the grocery purchases of 169,000 unique households across the United States, based on UPC records of all consumer packaged goods participants purchased from any outlet. (Figure 3). We compare against BRFSS rather than National Health and Nutrition Examination Survey (NHANES), since BRFSS is significantly larger than NHANES, it is remotely administered matching our study, and it has much better geographical coverage than NHANES and geographical comparisons are central to our study. Comparing our data to BRFSS on county level, we found strong correlations between the

4 amount of fresh fruits and vegetables (F&V) consumed (Figure 3a, R=0.63, p < 10−5) and body mass index (Figure 3b, R=0.78, p < 10−5). Comparing to USDA purchase data from the Nielsen Homescan Panel Survey we found that our app-based food logs were very highly correlated with previously published results (Figure 3c, R=0.88, p < 0.01) and that the absolute differences between food deserts and non-food deserts were stronger in the MFP data compared to Nielsen purchase data. See Supplementary Information for more details. These results demonstrate convergent validity and suggest that population-scale digital food logs can reproduce the basic dynamics of traditional, state-of-the-art measures, and they can do so at massive scale and comparatively low cost. 2.4 Statistical Analysis In this large-scale observational study, we used a matching-based ap- proach33, 34 to disentangle contributions of income, education, grocery access, and fast food access on food consumption. To estimate the impact of each of these factors, we divide all available zip codes into treatment and control groups based on a median split; that is, we estimate the difference in outcomes between matched above-median and below-median zip codes. We create matched pairs of zip codes by selecting a zip code in the control group that is closely matched to the zip code in the treatment group across all factors, except the treatment factor of interest. Since we repeat this matching process for each zip code in the treatment group, this approach estimates the Average Treatment Effect on the Treated (ATT). Through this process, we attempt to eliminate variation of plausible inﬂuences and to isolate the effect of interest. We repeat this process for each treatment of interest; for example for the results presented in Figure 4, we performed four matchings, one for each of income, education, grocery access and fast food access. For the sub- population experiments (Figure 5), we repeated the same method on the subset of the zip codes in which the majority of inhabitants were of a particular racial group. See Supplementary Information for details on the matching approach and detailed statistics that demonstrate that treatment and control groups were well-balanced on observed covariates after matching. We tested discriminant validity of our statistical approach by measuring the effect of null- treatments that should not have any impact on food consumption. We chose examples of null- treatments by selecting variables that had little correlation with study independent variables (income, education, grocery access, fast food access) and were plausibly unrelated to food consumption. This selection process lead to use of the fraction of countertop installers, electronics stores,

5 and waterprooﬁng services nearby as measured through Yelp. Applying our analysis pipeline to these null-treatments, we found that all of these null-treatments had zero effect on food consumption. This demonstrated that our statistical analysis approach did not produce measurements that it was not supposed to measure; that is, discriminant validity (Supplementary Figure e2).

6 3 Results Across all 9,822 U.S. codes, we found that high income, high education, high grocery access, and low fast food access were independently associated with higher consumption of fresh fruits and vegetables (F&V), lower consumption of fast food and soda, and lower prevalence of overweight or obese BMI levels (Figure 4).1 Speciﬁcally, in zip codes of above median grocery access participants logged 3.4% more F&V, 7.6% less fast food, 8.6% less soda and were 2.4% less likely to be overweight or obese (all P < 0.001). In zip codes of below median fast food access participants logged 5.3% more F&V, 6.2% less fast food, 15.7% less soda and were 1.5% less likely to be overweight or obese (all P < 0.001). In zip codes of above median education participants logged 9.2% more F&V, 8.5% less fast food, 10.6% less soda and were 13.1% less likely to be overweight or obese (all P < 0.001). Finally, in zip codes of above median household income (referred to as higher income below), participants logged 3.3% more F&V, 6.8% less fast food, 5.2% less soda (all P < 0.001), but had a 0.6% higher likelihood of being overweight or obese (P = 0.006). However, above median household income was associated with a 0.34% decrease in BMI (P < 0.001). Note that the reported effect size are based on comparing above and below median zip codes for any given factor. We found a general pattern of consistent but increased effect sizes when comparing top versus bottom quartiles (Supplementary Figure e1), suggesting a dose-response relationships across all considered variables (with the exception of the association between low fast food access and likelihood of overweight or obese BMI levels). We found that zip codes with high education levels compared to low education levels had the largest relative increases in F&V, fast food, and overweight or obese BMI levels. However, in terms of its impact on soda consumption, we found low fast food access to be associated with the largest relative decrease in soda consumption. In particular, high education zip codes had the largest relative decrease in reported soda consumption (15.7%). We separately repeated our data analyses within zip codes that were predominantly Black (3.7%), Hispanic (5.6%) and non-Hispanic White (78.4%) (Figure 5). Results within predominantly non-Hispanic White zip codes closely matched results within the overall population, since most zip codes in this study were predominantly White (78.4%), not unlike to the overall U.S. population (61.3%)35. However, restricting our analyses to predominantly Black and Hispanic zip

1The only exception to this pattern was a slight (0.34%) increase in overweight or obese BMI levels associated with income.

7 codes led to remarkably different findings. Specifically, within predominantly Black zip codes we found an impact of higher income in the inverse direction of the population average and towards low healthful food consumption, across across four out of four outcome variables, resulting in decreased F&V consumption (6.5%), increased fast food consumption (5.5%), and increased likelihood of overweight or obese BMI levels (8.1%). Higher income was also associated with increased soda consumption (11.8%) but was not statistically significant (P = 0.081). On the other hand, low fast food access and high education access were generally associated with increased diet health, with low fast food access correlating with the highest decrease in soda consumption (14.4%) and high education with the highest increase in fresh fruit and vegetable consumption (11.2%), although decreased fast food access was harmful to one of the outcome variables. Specif- ically, decreased fast food access was associated with a slight increase in fraction overweight or obese (3.1%). The only variable that had a positive effect on the health of all outcome variables was increased grocery access, which was associated with increased F&V consumption (10.2%), decreased fast food consumption (12.6%), decreased soda consumption (9.3%), and decreased likelihood of overweight or obese BMI levels (9.0%). In contrast, within predominantly Hispanic zip codes we found a significant effect of higher, above-median, income on higher F&V consumption (5.7%), but not across the remaining three outcome variables. Higher education zip codes had the most positive association with diet health across all variables, with the exception of soda consumption for which none of the factors had a significant impact. Specifically, higher education was associated with increased F&V consumption (8.9%), decreased fast food consumption (11.9%), and decreased likelihood of overweight or obese BMI levels (13.7%). Higher grocery access and decreased fast food access had similar effects as on the overall population for some outcome variables (i.e. similar associations with likelihood of overweight or obese BMI levels and fast food consumption). However, in some cases the magnitude of impact was higher (i.e. grocery access increased Hispanic F&V consumption by 7.4% which more than twice the increase of the overall population) and in others, unlike the overall population, there was no significant association (i.e. no significant relationship between fast food/grocery access on soda consumption, or between fast food access and F&V consumption). Few factors led to consistent improvements across all three subpopulations. Across all three groups, F&V consumption was strongly increased by high grocery access and education. Fast food

8 consumption was reduced by all potential intervention targets besides increased income. Soda consumption was most reduced by decreased fast food access for Black and White-majority zip codes, but was not impacted by any of the intervention targets within Hispanic zip codes. Lastly, overweight or obese BMI levels were, by far, most strongly reduced across all groups by increased education levels. 4 Discussion By analyzing 1.5 billion food intake logs and BMIs from 751,493 MyFitnessPal smartphone app users over seven years across 9,822 zip codes in relation to their demographic (education, ethnicity), socioeconomic (income), and food environment factors (grocery store and fast food access), we found higher access to grocery stores, lower access to fast food, higher income and education were independently associated with higher consumption of fresh fruits and vegetables, lower consumption of fast food and soda, and lower body mass index. This is consistent with previous studies36–38. As suggested in Larson’s review of neighborhood environment disparities in access to healthy foods in the U.S., residents who have better access to supermarkets and limited access to fast food restaurants have healthier diet and lower levels of obesity, and low-income, minority and rural neighborhoods often have limited access to supermarkets and higher availability of fast food restaurants36. Darmon39 and Hiza40 also found higher socioeconomic status (SES) groups had higher diet quality. While we report on differences between above/below median, we find significantly larger effect sizes for top vs bottom quartile (Supplementary Figure e1). This indicates that effects of food environment on diet are even more pronounced when considering larger differences in income, education, access, and fast food prevalence. Of note, we found high education zip codes had the largest relative decrease in BMI levels of overweight and obese by 13.1%, which suggests that intervention programs and policies that aimed at increase education levels of residents of low- income neighborhoods are essential in curbing the obesity trends. When we restricted our analyses to predominantly Black and Hispanic zip codes, we found the independent impacts of food access, income and education on food consumption and weight status varied significantly across Black, Hispanic and White populations. These findings suggest that tailored intervention strategies are needed based on neighborhood population distributions, assets and contexts.

9 Within predominantly Black zip codes, unlike other races, having higher income seems to be harmful as higher income Black zip codes had decreased fruits and vegetable consumption, increased fast food and soda consumption, and increased likelihood of overweight and obesity. The effect of higher education was low compared to grocery and fast food access. This could be explained by the “diminishing return hypothesis”, which suggests that Blacks receive fewer protective health beneﬁts from increases in SES than Whites41, 42. Research has documented that Blacks and Whites receive different levels of economic return for their location in the educational and occupational hierarchies43, and increases in SES (higher education and income) do not protect Blacks from increases in BMI in the same way as Whites44. A combination of factors, including neighborhood economic disadvantage45, 46, racial discrimination47, 48, and stress associated with educational attainment and mobility49, may prevent higher SES Blacks from achieving their fullest health potential relative to Whites44. Furthermore, in Sherman’s investigation of African American men’s perceptions of their food environment, the author showed that African American men perceived that fast food was their food choice and that they had no other healthy food options in or near their residences. This perception was shaped by their food environment, food marketing and advertising, the cost of eating healthy versus convenience, as well as their formative years23. In fact, it has been documented that African American and Hispanic children were disproportionately exposed to more fast food advertisements than White children50. Grier and Kumanyikas study showed that Black youth were exposed to as much as 70% more fast food and soda TV commercials than White youth. Black youth also had higher media use and spending patterns, which made them ideal target for marketers51. While there have been many public nutrition education campaigns and individual level interventions designed to change individual dietary behavior, these cannot be matched for the amount of advertising and marketing of unhealthy food52. Within predominantly Hispanic zip codes, higher income was not associated with reduce BMI levels, but higher education signiﬁcantly reduced BMI levels by 7%. High grocery access was the only factor associated with higher consumption of fresh fruits and vegetables, and high grocery and low fast food access had similar effects in reducing fast food and soda consumption and lowering overweight and obese BMI levels. The relationship between higher income and BMI could be partially explained by the “Hispanic health paradox” and “Hispanic health advantage”53.

10 The Hispanic health paradox suggests that even though the first-generation Hispanics have lower SES, they experience better health outcomes including lower prevalence of cardiovascular diseases, asthma, diabetes and cancer compared to those who were U.S.-born53–55. Hispanic health advantage suggests that Hispanics have lower rates of harmful health behaviors, such as smoking, which in turn positively influence other health outcomes compared to non-Hispanic Whites53, 56–58. Acculturation may also be another factor that influence Hispanics’ dietary behaviors. Previous research has shown that by adopting American culture, Hispanic immigrants engage in less healthy behaviors, which in turn put themselves at higher risk for chronic diseases53–55, 59–63. Obtaining higher education may increase the health literacy of Hispanics, which in turn could also influence their dietary behaviors, as well as preventive care use64. However, a recent study that examined atherosclerotic cardiovascular disease risk of a cohort of highly educated Hispanics found no evidence of the “Hispanic health paradox” in that Hispanics had the similar cardiovascular disease risk as non-Hispanic Whites with the same educational attainment, and this was likely due to acculturation61. Besides the individual health consequences, including increased risk of mortality, cardiovascular diseases, diabetes and certain cancers, obesity also brings direct and indirect economic consequences65. Cawley and Meyerhoefer estimated that obesity accounts for 21% of medical spending ($190 billion) in 200566, and if obesity trends continue to increase, obesity-related medical costs could rise by $48-66 billion a year67. More recent data indicated that the aggregate costs of obesity in the noninstitutionalized adult population of the US was as high as $315.8 billion in 2010, and the estimated extra annual medical care cost of an obese adult was $3,429 on average in 2013. The extra medical care costs are even higher for non-White populations ($4,086)68. Waters and Graf estimated chronic diseases caused obesity and overweight accounted for $480.7 billion in direct health care costs, plus additional $1.24 trillion in indirect costs attributed by lost productivity, which was 9.3 percent of the U.S. gross domestic product (GDP)69. The prevalence of overweight and obesity in the U.S., which was estimated to be 42.4% and about 107.6 million adults in 2017-201870. A 13.1% decrease by implementing effective education programs and policies (i.e., based on our estimate of above/below median effect size, see Figure 4), could potentially lead to more than $48.3 billions of annual health care cost savings71. According to the U.S. Department of Education, the 2019 presidents education budget for the entire U.S.

11 was $64 billion72, which is significantly less than the aggregate costs of obesity. The national postsecondary education budget, which include important funding supporting upward mobility such as the federal Pell grant, work study and student loans, was $24.2 billion in 202072. Based on our analysis, implementing effective education programs and policies could potentially lead to more than $48.3 billions of annual medical care cost savings in 2013 dollars. Taking inflation into account, this cost saving would be $53.6 billion in 2020, and could support 83.8% of the education budget and cover the entire postsecondary budget twice. Effective program and policy interventions could range from mandatory schooling policies, such as those promoting health and nutrition education at schools, to increasing educational quality, such as those aimed at promoting higher levels of education3, 73, 74. Similarly, having higher grocery access, and lower fast food access could potentially lead to $9.9 billions and $6.1 billions of annual health care cost savings today respectively (based on our estimates in Figure 4). While it is challenging to close the education and income gaps, establishing more grocery stores and limit fast food restaurants together could potentially save $16 billion. Pre- vious reviews suggested that government policies that addressing food affordability and purchase, such as the Healthy Food Financing Initiative (HFFI), increasing food stamp (SNAP) benefit and provide incentives to create healthy retail food environment have been effective in reducing food insecurity and dietary behaviors75–80. While several studies showed that the establishments of new supermarkets had little improvement in BMI81–83; however, the investments in the new supermarkets have improved economic opportunity and social cohesion84–86. Our results showed that the impact of having higher grocery store access increased fresh fruit and vegetable consumption and decreased fast food consumption 2-3 times than for Whites. Although previous literature has shown null effects of grocery access87, 88, these studies have focused on the general population, which is White-skewed. Therefore, policies and strategies in increasing grocery store access and decreasing fast food access could potentially be the most effective approaches in changing dietary habits among African Americans. Furthermore, having more grocery access and lower fast food access, in the food environment may work in synergistic ways that may lead to even lower obesity prevalence and obesity-related cost savings. This is demonstrated in a new study by Cantor et al. that HFFI boosted the effects of SNAP participation on improving food security and healthy food choices in food desserts89.

12 This synergy could be multiplied when combining with effective education programs that could potentially lower obesity prevalence further by increasing individuals’ SES (e.g., income and education)90, 91, health literacy and behaviors90–94, as well as sense of control and empowerment95. Due to the cross-sectional design of this observational study, we were not able to make any causal inferences between SES and food environment variables and dietary behavior and BMI. However, we used a matching-based approach to mimic a quasi-experimental design to disentangle the individual impact of income, education and food access on participants’ food consumption. Our analysis did not include other important variables such as gender and age, but we jointly considered the impacts of income, education and food environment access (grocery stores and fast food restaurants) on participants’ food consumption with consistent measures across the U.S., whereas many previously published studies examined one or a few at a time. We used individuals’ food loggings to estimate their consumption. Food loggings may not capture what individuals ac- tually ate. However, we conducted rigorous validations through comparisons with high quality and highly representative datasets which demonstrated high correlations to gold-standard approaches (Figure 3). Majority of the food environment studies used screeners, food frequency questionnaires or 24-hour recalls for dietary assessment, and very few used diaries7. Our participants logged their food intakes for an average of 197 days each. Our analysis was conducted at the zip code level. We used comparisons between above median and below-median zip codes across all dependent variables (SES and food environment) rather than using more speciﬁc cut-offs. In terms of zip code level income, our less wealthy zip codes are still relatively wealthy compared with average national income levels (Supplementary Table 4). Despite the ﬁne-grained spatial resolution, the analysis included a total of 9,822 zip codes, covering the vast majority of US counties. Two large neighborhood food environment studies included 20,897 participants from the Reasons for Geographic and Racial Differences in Stroke Study96 and 3,768 participants from the Framingham Heart Study26 respectively. Recently, Aiello et al. conducted an analysis using 1.6 billion food purchases and 1.1 billion medical prescriptions for the city of London; however, as the authors point out, food purchase data could not be used to construct individual dietary patterns97. We used individual’s food logging and BMI data generated from 1,164,926 participants using one of the most popular commercially available apps. Our study constitutes the largest United States nationwide study examining the impact of food environment

13 on diet to date. For this analysis, we also harnessed other large datasets such as Yelp to examine participants’ food environments. The usage of our combined dataset goes beyond food environment and diet studies. Other attributes such as participants physical activity levels and social networks can be utilized to study other built and social environment exposures and health outcomes. participants location tracking data not only can be used to derive environmental exposures such as air pollution and noise98, but also formulate personalized real-time intervention strategies99, 100. Con- sidering both the strengths and limitations of this study, more research is needed especially based on longitudinal study design and detailed individual level data to allow causal inference and precise interpretation of the results. 5 Conclusion We analyzed 2.3 billion food intake logs and BMIs from 1.2 million MyFitnessPal smartphone app participants over seven years across 9,822 zip codes in relation to education, race/ethnicity, income, and food environment access. Our analyses indicated that higher access to grocery stores, lower access to fast food, higher income and education were independently associated with higher consumption of fresh fruits and vegetables, lower consumption of fast food and soda, and lower likelihood of being overweight or obese. Policy targeted at improving food access, income and education may increase healthy eating. Potential interventions may need to be targeted to speciﬁc subpopulations for optimal effectiveness.

14 ARTICLE INFORMATION

Author Afﬁliations: Allen School of Computer Science & Engineering, University of Washing- ton (Tim Althoff); Department of Computer Science, Stanford University (Hamed Nilforoshan); Stanford Prevention Research Center, School of Medicine, Stanford University (Jenna Hua). De- partment of Computer Science, Stanford University (Jure Leskovec);

Author Contributions: Study concept and design: Althoff and Leskovec. Statistical analysis: Althoff and Nilforoshan. Interpretation of data: All authors. Drafting of the manuscript: All authors. Critical revision of the manuscript for important intellectual content: All authors.

Conﬂict of Interest Disclosure: None reported.

Funding/Support: T.A. and J.L. were supported by a National Institutes of Health (NIH) grant (U54 EB020405, Mobilize Center, NIH Big Data to Knowledge Center of Excellence). T.A. was supported by the SAP Stanford Graduate Fellowship, NSF grant IIS-1901386, and Bill Melinda Gates Foundation (INV-004841). H.N was supported by NSF REU #1659585 at the Stanford Cen- ter for the Study of Language and Information (CSLI). J.H is supported by Postdoctoral Fellowship in Cardiovascular Disease Prevention (T32) funded by the National Heart, Lung, and Blood Insti- tute (NHLBI) at the National Institutes of Health (NIH). J.L. was supported by DARPA under Nos. FA865018C7880 (ASED), N660011924033 (MCS); ARO under Nos. W911NF-16-1-0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC-1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions); Stanford Data Science Initiative, Wu Tsai Neurosciences In- stitute, Chan Zuckerberg Biohub, Amazon, Boeing, Chase, Docomo, Hitachi, Huawei, JD.com, NVIDIA, Dell.

Roles of Funder/Sponsor: The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Disclaimer: The content is solely the responsibility of the authors and does not necessarily repre- sent the ofﬁcial views of the NIH or sponsors.

Additional Contributions: We thank MyFitnessPal for donating the data for independent research and Mitchell Gordon for helping with data preparation.

15 Bibliography

1. Yach, D., Hawkes, C., Gould, C. L. & Hofman, K. J. The global burden of chronic diseases: overcoming impediments to prevention and control. JAMA 291, 2616–2622 (2004).

2. Beaglehole, R. et al. Priority actions for the non-communicable disease crisis. The Lancet 377, 1438–1447 (2011).

3. Fleischer, N. L., Roux, A. V. D., Alazraqui, M. & Spinelli, H. Social patterning of chronic disease risk factors in a latin american city. Journal of Urban Health 85, 923 (2008).

4. Sallis, J. F. & Glanz, K. The role of built environments in physical activity, eating, and obesity in childhood. The future of children 16, 89–108 (2006).

5. Fleischhacker, S. E., Evenson, K. R., Rodriguez, D. A. & Ammerman, A. S. A systematic review of fast food access studies. Obesity Reviews 12, e460–e471 (2011).

6. Odoms-Young, A., Singleton, C. R., Springﬁeld, S., McNabb, L. & Thompson, T. Retail environments as a venue for obesity prevention. Current obesity reports 5, 184–191 (2016).

7. Kirkpatrick, S. I. et al. Dietary assessment in food environment research: a systematic review. American Journal of Preventive Medicine 46, 94–102 (2014).

8. Caspi, C. E., Sorensen, G., Subramanian, S. & Kawachi, I. The local food environment and diet: a systematic review. Health & place 18, 1172–1187 (2012).

9. Feng, J., Glass, T. A., Curriero, F. C., Stewart, W. F. & Schwartz, B. S. The built environment and obesity: a systematic review of the epidemiologic evidence. Health & place 16, 175–190 (2010).

10. Cummins, S. & Macintyre, S. Food environments and obesity—neighbourhood or nation? International journal of epidemiology 35, 100–104 (2006).

11. Charreire, H. et al. Measuring the food environment using geographical information systems: a methodological review. Public health nutrition 13, 1773–1785 (2010).

16 12. Kelly, B., Flood, V. M. & Yeatman, H. Measuring local food environments: an overview of available methods and measures. Health & place 17, 1284–1293 (2011).

13. McKinnon, R. A., Reedy, J., Morrissette, M. A., Lytle, L. A. & Yaroch, A. L. Measures of the food environment: a compilation of the literature, 1990–2007. American Journal of Preventive Medicine 36, S124–S133 (2009).

14. Elinder, L. S. & Jansson, M. Obesogenic environments–aspects on measurement and indicators. Public Health Nutrition 12, 307–315 (2009).

15. Gittelsohn, J. & Sharma, S. Physical, consumer, and social aspects of measuring the food environment among diverse low-income populations. American Journal of Preventive Medicine 36, S161–S165 (2009).

16. Gustafson, A., Hankins, S. & Jilcott, S. Measures of the consumer food store environment: a systematic review of the evidence 2000–2011. Journal of community health 37, 897–911 (2012).

17. Lytle, L. A. Measuring the food environment: state of the science. American Journal of Preventive Medicine 36, S134–S144 (2009).

18. Odoms-Young, A. M., Zenk, S. & Mason, M. Measuring food availability and access in african-american communities: implications for intervention and policy. American Journal of Preventive Medicine 36, S145–S150 (2009).

19. Ohri-Vachaspati, P. & Leviton, L. C. Measuring food environments: a guide to available instruments. American Journal of Health Promotion 24, 410–426 (2010).

20. Sharkey, J. R. Measuring potential access to food stores and food-service places in rural areas in the us. American Journal of Preventive Medicine 36, S151–S155 (2009).

21. Kamphuis, C. B. et al. Environmental determinants of fruit and vegetable consumption among adults: a systematic review. British Journal of Nutrition 96, 620–635 (2006).

22. Lytle, L. A. & Sokol, R. L. Measures of the food environment: a systematic review of the ﬁeld, 2007–2015. Health & Place 44, 18–34 (2017).

17 23. Kumanyika, S. K. Environmental inﬂuences on childhood obesity: ethnic and cultural inﬂu- ences in context. Physiology & behavior 94, 61–70 (2008).

24. Sasson, C. et al. American heart association diabetes and cardiometabolic health summit: Summary and recommendations. Journal of the American Heart Association 7, e009271 (2018).

25. Mason, K. E., Pearce, N. & Cummins, S. Associations between fast food and physical activity environments and adiposity in mid-life: cross-sectional, observational evidence from uk biobank. The Lancet Public Health 3, e24–e33 (2018).

26. James, P., Seward, M. W., O’Malley, A. J., Subramanian, S. & Block, J. P. Changes in the food environment over time: examining 40 years of data in the framingham heart study. International Journal of Behavioral Nutrition and Physical Activity 14, 84 (2017).

27. Bureau, U. C. American community survey 5-year estimates (2014). URL http:// censusreporter.org. Accessed August 30, 2017.

28. Buajitti, E., Chiodo, S. & Rosella, L. C. Agreement between area-and individual-level income measures in a population-based cohort: Implications for population health research. SSM-population health 10, 100553 (2020).

29. Rhone, A. Food access research atlas documentation (2015). URL https: //www.ers.usda.gov/data-products/food-access-research-atlas/ documentation/. Accessed August 20, 2017.

30. CDC. Behavioral risk factor surveillance system survey questionnaire (2011). URL https: //www.cdc.gov/brfss/smart/smart 2011.htm. Accessed August 18, 2017.

31. CDC. Behavioral risk factor surveillance system survey questionnaire (2012). URL https: //www.cdc.gov/brfss/smart/smart 2012.htm. Accessed August 18, 2017.

32. Einav, L., Leibtag, E. S. & Nevo, A. On the accuracy of nielsen homescan data. Tech. Rep. (2008).

18 33. Austin, P. C. An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate Behavioral Research 46, 399–424 (2011).

34. Stuart, E. A. Matching methods for causal inference: A review and a look forward. Statistical Science: A review journal of the Institute of Mathematical Statistics 25, 1 (2010).

35. USCB. Quickfacts (2017). URL https://www.census.gov/quickfacts/fact/ table/US/PST045219.

36. Larson, N. I., Story, M. T. & Nelson, M. C. Neighborhood environments: disparities in access to healthy foods in the us. American Journal of Preventive Medicine 36, 74–81 (2009).

37. Mackenbach, J. D. et al. A systematic review on socioeconomic differences in the association between the food environment and dietary behaviors. Nutrients 11, 2215 (2019).

38. Meyer, K. A. et al. Sociodemographic differences in fast food price sensitivity. JAMA internal medicine 174, 434–442 (2014).

39. Darmon, N. & Drewnowski, A. Does social class predict diet quality? The American journal of clinical nutrition 87, 1107–1117 (2008).

40. Hiza, H. A., Casavale, K. O., Guenther, P. M. & Davis, C. A. Diet quality of americans differs by age, sex, race/ethnicity, income, and education level. Journal of the Academy of Nutrition and Dietetics 113, 297–306 (2013).

41. Farmer, M. M. & Ferraro, K. F. Are racial disparities in health conditional on socioeconomic status? Social Science & Medicine 60, 191–204 (2005).

42. Shuey, K. M. & Willson, A. E. Cumulative disadvantage and black-white disparities in life- course health trajectories. Research on Aging 30, 200–225 (2008).

43. Willson, A. E., Shuey, K. M. & Elder, G. H., Jr. Cumulative advantage processes as mech- anisms of inequality in life course health. American Journal of Sociology 112, 1886–1924 (2007).

19 44. Boen, C. The role of socioeconomic factors in black-white health inequities across the life course: Point-in-time measures, long-term exposures, and differential health returns. Social Science & Medicine 170, 63–76 (2016).

45. Do, D. P. et al. Does place explain racial health disparities? quantifying the contribution of residential context to the black/white health gap in the united states. Social Science & Medicine 67, 1258–1268 (2008).

46. Williams, D. R. & Collins, C. Us socioeconomic and racial differences in health: patterns and explanations. Annual Review of Sociology 21, 349–386 (1995).

47. Bratter, J. L. & Gorman, B. K. Is discrimination an equal opportunity risk? racial experiences, socioeconomic status, and health status among black and white adults. Journal of Health and Social Behavior 52, 365–382 (2011).

48. Williams, D. R., Neighbors, H. W. & Jackson, J. S. Racial/ethnic discrimination and health: ﬁndings from community studies. American Journal of Public Health 93, 200–208 (2003).

49. Pearson, J. A. Can’t buy me whiteness: New lessons from the titanic on race, ethnicity, and health. Du Bois Review: Social Science Research on Race 5, 27–47 (2008).

50. Powell, L. M., Zhao, Z. & Wang, Y. Food prices and fruit and vegetable consumption among young american adults. Health & place 15, 1064–1070 (2009).

51. Grier, S. A. & Kumanyika, S. K. The context for choice: health implications of targeted food and beverage marketing to african americans. American journal of public health 98, 1616–1629 (2008).

52. Crawford, P. B. Nutrition policies designed to change the food environment to improve diet and health of the population. In Nutrition Education: Strategies for Improving Nutrition and Healthy Eating in Individuals and Communities, vol. 92, 107–117 (Karger Publishers, 2020).

53. Lommel, L. L., Thompson, L., Chen, J.-L., Waters, C. & Carrico, A. Acculturation, inﬂam- mation, and self-rated health in mexican american immigrants. Journal of Immigrant and Minority Health 21, 1052–1060 (2019).

20 54. Bostean, G. Does selective migration explain the hispanic paradox? a comparative analysis of mexicans in the us and mexico. Journal of Immigrant and Minority Health 15, 624–635 (2013).

55. Ruiz, J. M., Steffen, P. & Smith, T. B. Hispanic mortality paradox: a systematic review and meta-analysis of the longitudinal literature. American Journal of Public Health 103, e52–e60 (2013).

56. Min, J. W., Rhee, S., Lee, S. E., Rhee, J. & Tran, T. Comparative analysis on determinants of self-rated health among non-hispanic white, hispanic, and asian american older adults. Journal of Immigrant and Minority Health 16, 365–372 (2014).

57. Kimbro, R. T., Gorman, B. K. & Schachter, A. Acculturation and self-rated health among latino and asian immigrants to the united states. Social Problems 59, 341–363 (2012).

58. Brewer, J. V. et al. Contributors to self-reported health in a racially and ethnically diverse population: focus on hispanics. Annals of Epidemiology 23, 19–24 (2013).

59. Ridker, P. M. C-reactive protein: a simple test to help predict risk of heart attack and stroke. Circulation 108, e81–e85 (2003).

60. Viruell-Fuentes, E. A. Beyond acculturation: immigration, discrimination, and health research among mexicans in the united states. Social Science & Medicine 65, 1524–1535 (2007).

61. Rodriguez, F. et al. Association of educational attainment and cardiovascular risk in hispanic individuals: Findings from the cooper center longitudinal study. JAMA Cardiology 4, 43–50 (2019).

62. Lutsey, P. L. et al. Associations of acculturation and socioeconomic status with subclinical cardiovascular disease in the multi-ethnic study of atherosclerosis. American Journal of Public Health 98, 1963–1970 (2008).

63. Koya, D. L. & Egede, L. E. Association between length of residence and cardiovascular disease risk factors among an ethnically diverse group of united states immigrants. Journal of General Internal Medicine 22, 841–846 (2007).

21 64. Daviglus, M. L., Pirzada, A. & Talavera, G. A. Cardiovascular disease risk factors in the hispanic/latino population: lessons from the hispanic community health study/study of latinos (hchs/sol). Progress in Cardiovascular Diseases 57, 230–236 (2014).

65. CDC. Adult obesity causes and consequences (2017). URL https://www.cdc.gov/ obesity/adult/causes.html.

66. Cawley, J. & Meyerhoefer, C. The medical care costs of obesity: an instrumental variables approach. Journal of Health Economics 31, 219–230 (2012).

67. Wang, Y. C., McPherson, K., Marsh, T., Gortmaker, S. L. & Brown, M. Health and economic burden of the projected obesity trends in the usa and the uk. The Lancet 378, 815–825 (2011).

68. Biener, A., Cawley, J. & Meyerhoefer, C. The high and rising costs of obesity to the us health care system (2017).

69. Waters, H. & Graf, M. America’s obesity crisis: The health and economic costs of excess weight. Available at https://www.milkeninstitute.org/reports/americas- obesity-crisis-health-and-economic-costs-excess-weight, note=Accessed March 05, 2020,.

70. CDC. Prevalence of obesity and severe obesity among adults: United states, 2017–2018 (2018). URL https://www.cdc.gov/nchs/products/databriefs/ db360.htm.

71. Census Quick Facts: Population. https://www.census.gov/quickfacts/fact/ table/US/PST045218. Accessed March 05, 2020.

72. of Education, U. D. Budget history tables (2020). URL https://www2.ed.gov/about/ overview/budget/tables.html?src=rt.

73. Cohen, A. K., Rai, M., Rehkopf, D. H. & Abrams, B. Educational attainment and obesity: a systematic review. Obesity Reviews 14, 989–1005 (2013).

74. Cohen, A. K. & Syme, S. L. Education: a missed opportunity for public health intervention. American Journal of Public Health 103, 997–1001 (2013).

22 75. Moran, A. et al. Financial incentives increase purchases of fruit and vegetables among lower- income households with children. Health Affairs 38, 1557–1566 (2019).

76. Moran, A. J. et al. Associations between governmental policies to improve the nutritional quality of supermarket purchases and individual, retailer, and community health outcomes: An integrative review. International journal of environmental research and public health 17, 7493 (2020).

77. Phipps, E. J. et al. Impact of a rewards-based incentive program on promoting fruit and vegetable purchases. American journal of public health 105, 166–172 (2015).

78. Steele-Adjognon, M. & Weatherspoon, D. Double up food bucks program effects on snap recipients’ fruit and vegetable purchases. BMC public health 17, 1–7 (2017).

79. Wilde, P., Klerman, J. A., Olsho, L. E. & Bartlett, S. Explaining the impact of usda’s healthy incentives pilot on different spending outcomes. Applied Economic Perspectives and Policy 38, 655–672 (2015).

80. Polacsek, M. et al. A supermarket double-dollar incentive program increases purchases of fresh fruits and vegetables among low-income families with children: the healthy double study. Journal of nutrition education and behavior 50, 217–228 (2018).

81. Cummins, S., Flint, E. & Matthews, S. A. New neighborhood grocery store increased aware- ness of food access but did not alter dietary habits or obesity. Health affairs 33, 283–291 (2014).

82. Dubowitz, T. et al. Changes in diet after introduction of a full service supermarket in a food desert. Health Affairs (Project Hope) 34, 1858 (2015).

83. Zhang, Y. T. et al. Is a reduction in distance to nearest supermarket associated with bmi change among type 2 diabetes patients? Health & place 40, 15–20 (2016).

84. Rogus, S., Athens, J., Cantor, J. & Elbel, B. Measuring micro-level effects of a new supermarket: do residents within 0.5 mile have improved dietary behaviors? Journal of the Academy of Nutrition and Dietetics 118, 1037–1046 (2018).

23 85. Chrisinger, B. A mixed-method assessment of a new supermarket in a food desert: contributions to everyday life and health. Journal of Urban Health 93, 425–437 (2016).

86. Giang, T., Karpyn, A., Laurison, H. B., Hillier, A. & Perry, R. D. Closing the grocery gap in underserved communities: the creation of the pennsylvania fresh food ﬁnancing initiative. Journal of Public Health Management and Practice 14, 272–279 (2008).

87. Jack, D. et al. Socio-economic status, neighbourhood food environments and consumption of fruits and vegetables in new york city. Public health nutrition 16, 1197–1205 (2013).

88. Strome, S., Johns, T., Scicchitano, M. J. & Shelnutt, K. Elements of access: the effects of food outlet proximity, transportation, and realized access on fresh fruit and vegetable consumption in food deserts. International quarterly of community health education 37, 61–70 (2016).

89. Cantor, J. et al. Snap participants improved food security and diet after a full-service supermarket opened in an urban food desert: Study examines impact grocery store opening had on food security and diet of supplemental nutrition assistance program participants living in an urban food desert. Health Affairs 39, 1386–1394 (2020).

90. Chandola, T., Clarke, P., Morris, J. & Blane, D. Pathways between education and health: a causal modelling approach. Journal of the Royal Statistical Society: Series A (Statistics in Society) 169, 337–359 (2006).

91. Cutler, D. M. & Lleras-Muney, A. Education and health: evaluating theories and evidence. Tech. Rep., National Bureau of Economic Research (2006).

92. Kenkel, D. S. Health behavior, health knowledge, and schooling. Journal of Political Econ- omy 99, 287–305 (1991).

93. Fletcher, J. M. & Frisvold, D. E. Higher education and health investments: does more schooling affect preventive health care use? Journal of Human Capital 3, 144–176 (2009).

94. Lleras-Muney, A. The relationship between education and adult mortality in the united states. The Review of Economic Studies 72, 189–221 (2005).

24 95. Mirowsky, J. Education, Social Status, and Health (Routledge, 2017).

96. Chen, M. et al. Association of community food environment and obesity among us adults: a geographical information system analysis. J Epidemiol Community Health 73, 148–155 (2019).

97. Aiello, L. M., Schifanella, R., Quercia, D. & Del Prete, L. Large-scale and high-resolution analysis of food purchases and health outcomes. EPJ Data Science 8, 14 (2019).

98. Khan, J., Ketzel, M., Kakosimos, K., Sørensen, M. & Jensen, S. S. Road trafﬁc air and noise pollution exposure assessment–a review of tools and techniques. Science of The Total Environment 634, 661–676 (2018).

99. Hardeman, W., Houghton, J., Lane, K., Jones, A. & Naughton, F. A systematic review of just-in-time adaptive interventions (jitais) to promote physical activity. International Journal of Behavioral Nutrition and Physical Activity 16, 31 (2019).

100. Kankanhalli, A., Saxena, M. & Wadhwa, B. Combined interventions for physical activity, sleep, and diet using smartphone apps: A scoping literature review. International journal of medical informatics 123, 54–67 (2019).

101. Rahkovsky, I. & Snyder, S. Food choices and store proximity (2015). URL https:// www.ers.usda.gov/webdocs/publications/45432/53943 err195.pdf?v= 42276/. Accessed December 31, 2019.

102. Wikipedia. List of fast food restaurant chains — Wikipedia, the free encyclope- dia. http://en.wikipedia.org/w/index.php?title=List%20of%20fast% 20food%20restaurant%20chains&oldid=933158460 (2017). [Online; accessed 01-September-2017].

103. Wikipedia. List of restaurant chains in the United States — Wikipedia, the free en- cyclopedia. http://en.wikipedia.org/w/index.php?title=List%20of% 20restaurant%20chains%20in%20the%20United%20States&oldid= 930612669 (2017). [Online; accessed 01-September-2017].

25 104. Hartlaub, P. Sweet! america’s top 10 brands of soda. http://www.nbcnews.com/id/ 42255151/ns/business-us business/t/sweet-americas-top-brands- soda (2017). [Online; accessed 01-September-2017].

105. CDC. Healthy weight, overweight, and obesity among u.s. adults (2018). URL https: //www.cdc.gov/nchs/data/nhsr/nhsr122-508.pdf.

106. CDC. Obesity and overweight (2016). URL https://www.cdc.gov/nchs/fastats/ obesity-overweight.htm.

107. Wilson, J. Americans getting older, new census ﬁgures show (2019). URL https://thehill.com/homenews/state-watch/449505-americans- getting-older-new-census-figures-show.

108. Worldbank. Population, female (% of total population) (2019). URL https:// data.worldbank.org/indicator/SP.POP.TOTL.FE.ZS.

109. USCB. Income, poverty and health insurance coverage in the united states: 2016 (2017). URL https://www.census.gov/newsroom/press-releases/2017/ income-povery.html.

110. Wilson, J. Census: More americans have college degrees than ever before (2017). URL https://thehill.com/homenews/state-watch/326995- census-more-americans-have-college-degrees-than-ever-before.

111. Din, A. Hud usps zip code crosswalk ﬁles (2017). URL https://www.huduser.gov/ portal/datasets/usps crosswalk.html. Accessed August 10, 2017.

112. Yelp. Yelp Fusion V3 API (2017). URL https://www.yelp.com/dataset/ documentation/main/. Accessed August 30, 2017.

113. Efron, B. & Tibshirani, R. J. An introduction to the bootstrap (CRC press, 1994).

114. Austin, P. C. & Small, D. S. The use of bootstrapping when using propensity-score matching without replacement: a simulation study. Statistics in medicine 33, 4306–4319 (2014).

115. Diamond, A. & Sekhon, J. S. Genetic matching for estimating causal effects (2012).

26 116. Teixeira, V., Voci, S. M., Mendes-Netto, R. S. & da Silva, D. G. The relative validity of a food record using the smartphone application myﬁtnesspal. Nutrition & Dietetics 75, 219– 225 (2018).

117. Grifﬁths, C., Harnack, L. & Pereira, M. A. Assessment of the accuracy of nutrient calcula- tions of ﬁve popular nutrition tracking applications. Public Health Nutrition 21, 1495–1502 (2018).

118. Chen, J., Berkman, W., Bardouh, M., Ng, C. Y. K. & Allman-Farinelli, M. The use of a food logging app in the naturalistic setting fails to provide accurate measurements of nutrients and poses usability challenges. Nutrition 57, 208–216 (2019).

119. Nguyen, Q. et al. Social media indicators of the food environment and state health outcomes. Public Health 148, 120–128 (2017).

120. Gomez-Lopez, I. N. et al. Using social media to identify sources of healthy food in urban neighborhoods. Journal of Urban Health 94, 429–436 (2017).

121. Glass, G. V. Primary, secondary, and meta-analysis of research. Educational Researcher 5, 3–8 (1976).

27 Figure 1: Number of participants in our study across U.S. counties. A choropleth showing the number of participants in each U.S. county. This country-wide observational study included 1,164,926 participants across 9,822 U.S. zip codes that collectively logged 2.3 billion food entries for an average of 197 days each. Most U.S. counties are represented by at least 30 participants in our dataset. This study constitutes the largest nationwide study examining the impact of food environment on diet to date, with 300 times more participants and 4 times more person years of tracking than the Framingham Heart Study26.

28 Figure 2: Dietary Consumption and Body Weight across U.S. Counties. A set of choropleths showing the main study outcomes of the number of entries that are classified as fresh fruit and vegetables, fast food, and soda consumption as well as the fraction of overweight/obese participants 29 across the USA by counties with more than 30 participants. We observe that food consumption healthfulness varies significantly across counties in the United States. Figure 3: This studies’ smartphone-based food logs correlate with previous, representative survey measures and purchase data. a, Fraction of fresh fruits and vegetables logged is correlated with BRFSS survey data30 (R=0.63, p < 10−5; Methods). b, Body Mass Index of smartphone cohort is correlated with BRFSS survey data31 (R=0.78, p < 10−5; Methods). c, Digital food logs replicate previous findings of relative consumption differences in low-income, low-access food deserts based on Nielsen purchase data101 (R=0.88, p < 0.01; Methods). These results demonstrate that smartphone-based food logs are highly correlated with existing, gold-standard survey measures and purchase data.

30 a. b.

Fresh Fruits and Vegetables Consumption Fast Food Consumption

17.5 17.5

15.0 15.0

12.5 12.5

10.0 10.0

7.5 7.5

5.0 5.0

2.5 2.5

0.0 0.0 Relative Increase in Consumption (in %) 2.5 Relative Decrease in Consumption (in %) 2.5

High Income High Grocery High Income High Grocery High Education Low Fast Food High Education Low Fast Food

c. d.

Soda Consumption % Overweight or Obese

17.5 17.5

15.0 15.0

12.5 12.5

10.0 10.0

7.5 7.5

5.0 5.0

2.5 2.5

0.0 0.0

Relative Decrease in Consumption (in %) 2.5 2.5 Relative Decrease in Fraction Overweight or Obese (in %) High Income High Grocery High Income High Grocery High Education Low Fast Food High Education Low Fast Food Figure 4: The impact of income, education, grocery store and fast food access on food consumption and weight status. Independent contributions of high income (median family income higher than or equal to $70,241), higher education (fraction of population with college education 29.8% or higher), high grocery access (fraction of population that is closer than 0.5 miles from nearest grocery store is greater than or equal to than 20.3%), and low fast food access (less than or equal to 5.0% of all businesses are fast-food chains) on relative change in consumption of a, fresh fruits and vegetables, b, fast food, c, soda, and d, relative change in percent overweight or obese (BMI 25+). Cut points correspond to median values. Estimates are based on matching experiments controlling for all but the one treatment variable in focus (Methods). Error bars correspond to bootstrapped 95% conﬁdence intervals (Supplementary Methods). While the most impactful factors vary across outcomes, only higher education was associated with a sizeable decrease of 13.1% in overweight and obese weight status. 31 a. b.

20 20 White White Black Black Hispanic Hispanic

10 10

0 0

−10 −10

−20 −20 Relative Decrease in Fast Food Consumption (in %) Relative Increase in Fresh Fruits and Vegetables Consumption (in %) −30 −30

High Income High Income Low Fast Food High Grocery Low Fast Food High Grocery c. High Education d. High Education

20 20 White White Black Black Hispanic Hispanic

10 10

0 0

−10 −10

−20 −20 Relative Decrease in Non-Diet Soda Consumption (in %) Relative Decrease in Fraction Overweight or Obese (in %)

−30 −30

High Income High Income Low Fast Food High Grocery Low Fast Food High Grocery High Education High Education Figure 5: Effects on food consumption and weight status disaggregated by predominantly Black, Hispanic, and non-Hispanic White zip codes (i.e., 50% or more). Independent contributions of high income (median family income higher than or equal to $70,241), higher education (fraction of population with college education 29.8% or higher), high grocery access (fraction of population that is closer than 0.5 miles from nearest grocery store is greater than or equal to than 20.3%), and low fast food access (less than or equal to 5.0% of all businesses are fast-food chains) on relative change in consumption of a, fresh fruits and vegetables, b, fast food, c, soda, and d, relative change in percent overweight or obese (BMI > 25). Cut points correspond to median values. Estimates are based on matching experiments controlling for all but one treatment variable (Methods). Error bars correspond to bootstrapped 95% conﬁdence intervals (Methods). We observe signiﬁcant differences across Black, Hispanic, and non-Hispanic White zip codes in terms how food consumption is affected by factors of income, education, fast food and grocery access. 32 Supplemental Materials • Supplemental Methods

• Supplemental Figures

• Supplemental Tables

List of Figures e1 Matching experiments using quartile instead of median split. Note the consistent but increased effect sizes compared to Figure 4...... 13 e2 Demonstration of discriminant validity of statistical approach. We measured the effect of null-treatments that should not have any impact on food consumption. We chose examples of null-treatments by selecting variables that had little correlation with study independent variables (income, education, grocery access, fast food access) and were plausibly unrelated to food consumption. This selection process lead to use of the fraction of countertop installers, electronics stores, and waterprooﬁng services nearby as measured through Yelp. Applying our analysis pipeline to these null-treatments, we found that all of these effect estimates were close to zero. This demonstrated that our statistical analysis approach did not produce measurements that it was not supposed to measure; that is, discriminant validity...... 14 List of Tables e1 Demographic statistics for our study compared with nationally representative survey data. (*) indicates statistics calculated at the zip code level...... 4 e2 Outcome measures calculated at the zip code level for the 9,822 zip codes in our study, spanning 1,164,926 participants...... 6 e3 Effect sizes of all top/bottom half matching experiments (Fig. 4)...... 10 e4 Effect sizes of all race-speciﬁc top/bottom half matching experiments (Fig. 5). . . . 11 e5 Effect sizes of all top/bottom quartiles matching experiments (Fig. e1)...... 12 e6 USA fast food restaurants table used to classify participant food entries as fast food102. A list of popular pizza chains from the USA was appended to the list103.. 15

1 e7 USA soda list used to classify participant food entries as sugary sodas. The list was constructed using a list of America’s best-selling brands of Soda104, in addition to the generic terms such “Root Beer” and “Coca”...... 16 e8 Summary of High Income (MedianFamilyIncome > Median) matching experiment 17 e9 Summary of High Grocery (Grocery Access > Median) matching experiment . . . 18 e10 Summary of High Education (% College Degrees > Median) matching experiment 19 e11 Summary of Low Fast Food (% Yelp Fast Food < Median) matching experiment . 20 e12 Summary of High Income (MedianFamilyIncome > 75th Percentile) matching experiment ...... 21 e13 Summary of High Grocery (Grocery Access > 75th Percentile) matching experiment 22 e14 Summary of High Education (% College Degrees > 75th Percentile) matching experiment. Note: Treatment samples dropped due to 0.3 STD caliper used to ensure 0.25 SMD balancing constraint...... 23 e15 Summary of Low Fast Food (% Yelp Fast Food < 25th Percentile) matching experiment. Note: Treatment samples dropped due to 1.5 STD caliper used to ensure 0.25 SMD balancing constraint...... 24 e16 Summary of Low Countertop Installation Services (% Yelp Countertop Installers < Median) matching experiment ...... 25 e17 Summary of Low Electronics Stores (% Yelp Electronics Stores < Median) matching experiment ...... 26 e18 Summary of Low Waterprooﬁng Services (% Yelp Waterprooﬁng Services < Me- dian) matching experiment ...... 27 e19 Summary of Black-majority Zip Code High Income (MedianFamilyIncome > Me- dian) matching experiment ...... 28 e20 Summary of Hispanic-majority Zip Code High Income (MedianFamilyIncome > Median) matching experiment ...... 29 e21 Summary of White-majority Zip Code High Income (MedianFamilyIncome > Me- dian) matching experiment ...... 30 e22 Summary of Black-majority Zip Code High Grocery (Grocery Access > Median) matching experiment ...... 31

2 e23 Summary of Hispanic-majority Zip Code High Grocery (Grocery Access > Me- dian) matching experiment ...... 32 e24 Summary of White-majority Zip Code High Grocery (Grocery Access > Median) matching experiment ...... 33 e25 Summary of Black-majority Zip Code High Education (% College Degrees > Me- dian) matching experiment ...... 34 e26 Summary of Hispanic-majority Zip Code High Education (% College Degrees > Median) matching experiment ...... 35 e27 Summary of White-majority Zip Code High Education (% College Degrees > Me- dian) matching experiment. Note: Treatment samples dropped due to 2.1 STD caliper used to ensure 0.25 SMD balancing constraint...... 36 e28 Summary of Black-majority Zip Code Low Fast Food (% Yelp Fast Food < Me- dian) matching experiment ...... 37 e29 Summary of Hispanic-majority Zip Code Low Fast Food (% Yelp Fast Food < Median) matching experiment ...... 38 e30 Summary of White-majority Zip Code Low Fast Food (% Yelp Fast Food < Me- dian) matching experiment ...... 39

3 Supplemental Methods Details on food environment measures We obtained data on grocery store access (fraction of population that is more than 0.5 miles away from grocery store) and food desert status from the USDA Food Access Research Atlas29. A census tract is considered a food desert by the USDA if it is both low-income (defined by Department of Treasury’s New Markets Tax Credit program) and low-access, meaning at least 500 people or 30 percent of residents live more than 0.5 miles from a supermarket in urban areas (10 miles in rural areas)101. Although the USDA uses different thresholds for urban and rural areas (0.5 miles and 10 miles respectively), we found that even in non-Urban zipcodes (defined as USDA rural-urban con- tinuum RUCA scores of 4 or higher), the fraction of that population that is farther than 0.5 miles from grocery stores had the highest correlation to Fruit Vegetable consumption (R=-0.19), compared to 1 miles (R=-0.13), 10 miles (R=-0.07), and 20 miles (R=0.02). This suggests that the fraction of the population farther than 0.5 miles from a grocery store has the strongest relationship with healthy food consumption, even in non-Urban zipcodes. Hence, we decided used 0.5 miles distance as a standard measure of grocery access for both rural and urban zip codes, contrary to the USDA definition. We subsequently sanity checked for any downstream confounding of urbanicity in our primary matching experiment of above/below median grocery access, and found a negligible difference (Standardized Mean Difference of 0.18) in urbanicity between control and treatment, suggesting that the effect size was not due to grocery store distance functioning as a proxy for urbanicity, but rather directly due to differential grocery access. We aggregated these data from a census tract level to a zip code level using USPS Crosswalk data, which provides a list of all census tracts which overlap with a single zip code111. We related these data on census tract level to the zip code level by taking the weighted average of each census

Table e1: Demographic statistics for our study compared with nationally representative survey data. (*) indicates statistics calculated at the zip code level.

Overweight Median Source BMI Obese Median Age Gender College∗ Ethnicity∗ or Obese Family Income∗

Our Study 28.5 67.8% 32.8% 36 74% Female $ 76,563 33.7% 68.3% White Nat.l Avg. 29.4105 71.6%106 39.8%106 38.2107 50.5% Female108 $ 59,039109 33.4%110 61.3% White35

4 tract food environment measure (both grocery store access and food desert status), weighted by the number of people in the tract111. For instance, if zip code A overlapped with Census Tract A (2500 people, food desert) and Census Tract B (7500 people, not a food desert), the food desert measure of zip code A would be estimated as 25%. We defined the binary threshold for food desert, used in Figure 3, as 50% or higher. We measured fast food access through the fraction of restaurants in a zip-code that are fast food restaurants. Data on local restaurants and businesses were obtained through the Yelp API112. For each zip code, we consider up to 1000 restaurant businesses that are nearest to the zip code center up to a distance of 40km (67.8% of zip code queries resulted in 1000 restaurant businesses within 40km; Yelp API results are restricted to 1000 results). This resulted in a varying sample radius depending on urbanicity. For example, Urban zipcodes (RUCA code of 1) had an average effective centroid size of 15 miles, which we calculated by taking the distance from the zipcode center to the furthest restaurant returned by Yelp. We further used Yelp-based environment variables that we expected not to influence food consumption, such as the availability of waterproofing services, countertop installers, or electronic stores, as null experiments to demonstrate discriminant validity of our statistical analysis pipeline (see Supplementary Figure e2). Details on outcome measure (food consumption and weight status) We used 2.3 billion food intake logs by 1,164,926 U.S. participants of the MyFitnessPal (MFP) smartphone application to quantify food consumption across 9,822 zip codes. During the observation period from January 1, 2010 to November 15, 2016, the average participant logged 9.30 entries into their digital food journal per day. The average participant used the app for 197 days. All participants in this sample used the app for at least 10 days. Each entry contains a separate food component (e.g., banana, yogurt, hamburger, ...). We classified all entries into three categories: fresh fruits and vegetables (F&V; through a proprietary classifier by MFP), fast food (if the description contained the name of a fast food chain listed in Supplementary Information Table e6, and sugary (non-diet) soda (if the description contained the name of a soda drink listed in Supplementary Table e7 but did not contain “diet”, “lite”, “light”, or “zero”). In all cases, descriptions, as well fast food and soda drink keywords, were normalized by lower-casing and removing punctuation. We then calculated the average number of food entries logged per participant, per day, for each of the F&V, fast food, and soda categories (e.g. average number of F&V logged per participant per day), excluding days in

5 Table e2: Outcome measures calculated at the zip code level for the 9,822 zip codes in our study, spanning 1,164,926 participants.

F&V Entries Soda Entries Fast Food Entries % Overweight # Participants BMI % Obese per Day per Day per Day or Obese

mean 118.6 0.61 0.04 0.39 28.8 69 35 median 90 0.60 0.04 0.38 28.8 70 35 std 91.5 0.11 0.02 0.10 1.6 10 11 min 30 0.25 0.001 0.12 23.2 17 2 max 1262 1.29 0.17 0.91 36.9 100 80

which the participant was inactive (i.e., consistently did not log anything). Finally, we aggregated these participant-level measures to the zip code level by taking the mean of each category’s measure for all participants in each zip code. We further used body-mass index (BMI) health in each zip code as a weight status outcome. We operationalize BMI health as the fraction of participants in a zip code which are overweight or obese (BMI > 25). BMI was self-reported by participants of the smartphone application (99.92% of participants did report BMI). Supplementary Table e2 shows basic summary statistics for the outcome measures used in this study. In our statistical analyses, we compared two sets of zip codes that differ in a dimension of interest (e.g., grocery store access access) as treatment and control group and use the relative change in F&V consumption, fast food consumption, soda consumption, and BMI health of the treated group relative to the control group. To generate confidence intervals, as well as to compute p-values to test for statistical significance of changes in outcome, we use non-parametric bootstrap resampling113. Specifically, we follow the method proposed by Austin and Small,114 which is to draw bootstrap samples post-matching from the matched pairs in the propensity-score-matched sample after the Genetic Matching stage115. We confirmed the validity of this method empirically by also calculating t-tests for each experiment, which gave qualitatively similar results. Data Validation We find that our study population has significant overlap with the U.S. national population (Supplementary Table e1) but is skewed towards women and higher income. We demonstrate that food consumption measured based on this population are highly correlated with state-of-the-art measures (Figure 3, Section 2.3). Smartphone apps such as MyFitnessPal fea-

6 ture large databases with nutritional information and can be used to track one’s diet over time. Previous studies have compared app-reported diet measures to traditional measures including 24 hour dietary recalls and food composition tables. These studies found that both measures tend to be highly correlated116, 117, but that app-reported measures tend to underestimate certain macro- and micronutrients116, 117, especially in populations that were previously unfamiliar with the smartphone applications118. In contrast, this study leverages a convenience sample of existing participants of the smartphone app MyFitnessPal. Yelp data has been used in measures of food environment119 and a study in Detroit found Yelp data to be more accurate than commercially- available databases such as Reference USA120. This study uses a combination of MFP data to capture food consumption, Yelp, and USDA data to capture food environment, and Census data to capture basic demographics. As a preliminary, basic test, we investigated correlations between the Mexican food consumption, the fraction of Mexican restaurants, and the fraction of Hispanics in the population, on a zip code level. We found that Mexican food consumption (entries labeled as Mexican food by a proprietary MFP classifier, logged per participant, per day) was correlated with the fraction of Mexican restaurants (R=0.72; < 10−4) and the fraction of Hispanics in the population (R=0.54; P < 10−4). Further, the fraction of Mexican restaurants was correlated with the fraction of Hispanics in the population as well (R=0.51; P < 10−4). Details on reproducing state-of-the-art measures using population-scale digital food logs We used the latest survey data from BRFSS30, 31 available at the county-level. Specifically, we used variables FV5SRV from BRFSS 2011 representing the the faction of people eating five or more servings of fresh fruit and vegetables30, and BMI5 from BRFSS 2012 representing body mass index31. We compare against BRFSS rather than National Health and Nutrition Examination Sur- vey (NHANES), since BRFSS is significantly larger than NHANES, it is remotely administered matching our study, and it has much better geographical coverage than NHANES and geographical comparisons are central to our study. Comparing to BRFSS on a county level, the average number of F&V logged per day (MFP) was correlated with the fraction of respondents that report consumptions of at least five servings of F&V per day (Figure 3a, R=0.63, p < 10−5). Further, average county-level BMI was strongly correlated as well (Figure 3b, R=0.78, p < 10−5). We further compared to published results by the USDA101, which used data from the 2010 Nielsen Homes- can Panel Survey that captured household food purchases for in-home consumption (but did not

7 capture restaurants and fast food purchases). We attempted to reproduce published findings on the differences in low-income, low-access communities (food deserts) compared to non-low-income, non-low-access communities101 across categories of fruit, vegetable, sweets, red meat, fish/poultry, milk products, diet drinks, and non-diet drinks (Table 4 in Rahkovsky and Snyder101). We used proprietary MFP classifiers to categorize foods logged into these categories. We found that our app-based food logs were very highly correlated with previously published results (Figure 3c, R=0.88, p < 0.01) and that the absolute differences between food deserts and non-food deserts were stronger in the MFP data compared to Nielsen purchase data. In total, these results demonstrate convergent validity. Specifically, our results suggest that population-scale digital food logs can reproduce the basic dynamics of traditional, state-of-the-art measures, and they can do so at massive scale and comparatively low cost. Details on Matching Approach Specifically, we use a one-to-one Genetic Matching approach115, with replacement, and use the mean of the Standardized Mean Difference (SMD) between treatment and control groups, across all matched variables, as the Genetic Matching balance metric in order to maximize balance (overlap) between the treated and the control units. Some definitions of SMD use the standard variation in the overall population before matching33. However, we choose the standard deviation in the control group post-matching, which typically is much smaller and therefore gives more conservative estimates of balance between treated and control units121. After matching, we evaluated the quality of balance between the treated and the control units by the Standardized Mean Difference across each of the variables that were controlled for and included in the matching process. A good balance between treated and control groups was defined as a Standardized Mean Difference (SMD) of less than 0.25 standard deviations34 across each variable. By default, we do not enforce a caliper in order to minimize bias in matching process, although in rare cases in which a good balance was not achieved, a caliper was enforced, starting at 2.5 standard deviations between matched and controlled units, and decreased by 0.1 until the matched and control groups had a SMD smaller than 0.25 across all matched variables. For the vast majority of matching experiments the SMD across all matched variables was well below 0.25, with a mean of 0.040 and median of 0.016 for the four overall population matching experiments. The SMD for the race-majority zipcode experiments was slightly higher, but still very significantly below 0.25 across all 12 experiments, with a mean of 0.055 and median of 0.036.

8 Thus, no caliper was necessary to ensure a good balance, with the exception of one out of the 12 of sub-population experiments (White, High Education). Detailed balancing statistics for each of the matches are available in the Appendix (Tables e8a-e30b), as well as a supplementary matching experiment in which a top/bottom quartile split was used instead of a median split (Figure e1). Details on obesity prevention related cost savings It was estimated that the aggregate costs of obesity in the non-institutionalized adult population of the U.S. was as high as $315.8 billion in 2010, and the estimated extra annual medical care cost of an obese adult was $3,429 on average in 2013. The extra medical care costs are even higher for non-White populations ($4,086)68. The prevalence of overweight and obesity among US adults was 42.4% and about 107.6 million adults in 2017-201870. A 13.1% decrease by implementing effective education programs and policies (i.e., based on our estimate of above/below median effect size, see Figure 4), could potentially lead to more than $48.3 billions of annual health care cost savings71. To produce this estimate we need to assume that the Average Treatment Effect on the Treated (ATT; estimated through our matching procedure) can be generalized to the overall U.S. population. According to the U.S. Department of Education, the 2019 president’s education budget for the entire U.S. was $64 billion72, which is signiﬁcantly less than the aggregate costs of obesity. The national postsecondary education budget, which include important funding supporting upward mobility such as the federal Pell grant, work study and student loans, was $24.2 billion in 202072. Based on our analysis, implementing effective education programs and policies could potentially lead to more than $48.3 billions of annual medical care cost savings in 2013 dollars. Taking inﬂation into account, this cost saving would be $53.6 billion in 2020, and could support 83.8% of the education budget and cover the entire postsecondary budget twice. Effective program and policy interventions could range from mandatory schooling policies, such as those promoting health and nutrition education at schools, to increasing educational quality, such as those aimed at promoting higher levels of education3, 73, 74. Similarly, having higher grocery access (2.4% decrease in overweight and obesity), and lower fast food access (1.5% decrease in overweight and obesity) could potentially lead to $9.9 billions and $6.1 billions of annual health care cost savings today respectively (based on our estimates in Figure 4).

9 Table e3: Effect sizes of all top/bottom half matching experiments (Fig. 4).

Treatment Outcome % Change Trt. Mean Ctrl. Mean P (bootstrapping)

High Income Fresh F&V Consumption 3.265 0.631 0.652 < 0.001 High Income Fast Food Consumption -6.772 0.367 0.342 < 0.001 High Income Soda Consumption -5.219 0.036 0.034 < 0.001 High Income BMI -0.335 28.077 27.983 < 0.001 High Income % Overweight or Obese 0.643 0.647 0.651 0.006 High Education Fresh F&V Consumption 9.180 0.602 0.657 < 0.001 High Education Fast Food Consumption -8.457 0.371 0.340 < 0.001 High Education Soda Consumption -10.600 0.038 0.034 < 0.001 High Education BMI -5.053 29.266 27.787 < 0.001 High Education % Overweight or Obese -13.100 0.734 0.638 < 0.001 Low Fast Food Fresh F&V Consumption 5.311 0.612 0.645 < 0.001 Low Fast Food Fast Food Consumption -6.176 0.370 0.347 < 0.001 Low Fast Food Soda Consumption -15.725 0.039 0.033 < 0.001 Low Fast Food BMI -0.335 28.467 28.371 < 0.001 Low Fast Food % Overweight or Obese -1.474 0.679 0.669 < 0.001 High Grocery Fresh F&V Consumption 3.437 0.606 0.627 < 0.001 High Grocery Fast Food Consumption -7.581 0.390 0.361 < 0.001 High Grocery Soda Consumption -8.567 0.039 0.036 < 0.001 High Grocery BMI -0.698 28.839 28.638 < 0.001 High Grocery % Overweight or Obese -2.437 0.699 0.682 < 0.001 Low Countertop Svc. Fresh F&V Consumption -0.177 0.612 0.610 0.289 Low Countertop Svc. Fast Food Consumption -0.190 0.384 0.384 0.337 Low Countertop Svc. Soda Consumption 0.467 0.040 0.040 0.270 Low Countertop Svc. BMI 0.102 28.710 28.740 0.082 Low Countertop Svc. % Overweight or Obese 0.012 0.690 0.690 0.499 Low Electronics Stores Fresh F&V Consumption -0.681 0.609 0.605 0.011 Low Electronics Stores Fast Food Consumption -0.482 0.387 0.386 0.118 Low Electronics Stores Soda Consumption -0.406 0.040 0.040 0.278 Low Electronics Stores BMI 0.006 28.906 28.908 0.467 Low Electronics Stores % Overweight or Obese 0.018 0.701 0.701 0.467 Low Waterproofing Svc. Fresh F&V Consumption -1.607 0.614 0.604 < 0.001 Low Waterproofing Svc. Fast Food Consumption -1.319 0.389 0.384 < 0.001 Low Waterproofing Svc. Soda Consumption 0.600 0.040 0.040 0.222 Low Waterproofing Svc. BMI 0.046 28.786 28.799 0.274 Low Waterproofing Svc. % Overweight or Obese 0.239 0.694 0.695 0.133

10 Table e4: Effect sizes of all race-speciﬁc top/bottom half matching experiments (Fig. 5).

Race Treatment Outcome % Change Trt. Mean Ctrl. Mean P-value (bootstrapping)

Black High Income Fresh F&V Consumption -6.497 0.607 0.567 0.004 Black High Income Fast Food Consumption 5.464 0.403 0.425 0.015 Black High Income Soda Consumption 11.781 0.033 0.037 0.081 Black High Income BMI 3.695 29.834 30.936 < 0.001 Black High Income % Overweight or Obese 8.101 0.750 0.811 < 0.001 Black High Education Fresh F&V Consumption 11.243 0.558 0.620 < 0.001 Black High Education Fast Food Consumption -7.613 0.424 0.391 0.002 Black High Education Soda Consumption 6.088 0.033 0.035 0.150 Black High Education BMI -5.512 31.362 29.634 < 0.001 Black High Education % Overweight or Obese -11.481 0.829 0.734 < 0.001 Black Low Fast Food Fresh F&V Consumption 7.024 0.543 0.582 < 0.001 Black Low Fast Food Fast Food Consumption -12.043 0.473 0.416 < 0.001 Black Low Fast Food Soda Consumption -14.421 0.040 0.034 < 0.001 Black Low Fast Food BMI 0.506 30.776 30.931 0.153 Black Low Fast Food % Overweight or Obese 3.062 0.775 0.799 0.001 Black High Grocery Fresh F&V Consumption 10.230 0.527 0.581 < 0.001 Black High Grocery Fast Food Consumption -12.642 0.478 0.418 < 0.001 Black High Grocery Soda Consumption -9.300 0.039 0.036 0.002 Black High Grocery BMI -3.795 31.966 30.753 < 0.001 Black High Grocery % Overweight or Obese -8.960 0.861 0.783 < 0.001 Hispanic High Income Fresh F&V Consumption 5.706 0.556 0.588 0.012 Hispanic High Income Fast Food Consumption -3.314 0.393 0.380 0.140 Hispanic High Income Soda Consumption 0.663 0.036 0.036 0.429 Hispanic High Income BMI 0.289 28.997 29.080 0.280 Hispanic High Income % Overweight or Obese -0.039 0.722 0.722 0.492 Hispanic High Education Fresh F&V Consumption 8.859 0.575 0.626 < 0.001 Hispanic High Education Fast Food Consumption -11.902 0.385 0.339 < 0.001 Hispanic High Education Soda Consumption -8.530 0.034 0.031 0.061 Hispanic High Education BMI -5.949 29.738 27.969 < 0.001 Hispanic High Education % Overweight or Obese -13.697 0.758 0.654 < 0.001 Hispanic Low Fast Food Fresh F&V Consumption 1.501 0.562 0.570 0.083 Hispanic Low Fast Food Fast Food Consumption -5.920 0.423 0.398 < 0.001 Hispanic Low Fast Food Soda Consumption -0.538 0.038 0.038 0.433 Hispanic Low Fast Food BMI -0.172 29.750 29.699 0.291 Hispanic Low Fast Food % Overweight or Obese -1.800 0.763 0.750 0.004 Hispanic High Grocery Fresh F&V Consumption 7.351 0.525 0.563 < 0.001 Hispanic High Grocery Fast Food Consumption -7.212 0.443 0.411 < 0.001 Hispanic High Grocery Soda Consumption -1.956 0.040 0.039 0.178 Hispanic High Grocery BMI -1.518 30.358 29.898 < 0.001 Hispanic High Grocery % Overweight or Obese -3.492 0.791 0.763 < 0.001 White High Income Fresh F&V Consumption 2.182 0.643 0.657 < 0.001 White High Income Fast Food Consumption -5.063 0.359 0.341 < 0.001 White High Income Soda Consumption -5.512 0.037 0.035 < 0.001 White High Income BMI 0.486 27.759 27.894 < 0.001 White High Income % Overweight or Obese 3.261 0.625 0.646 < 0.001 White High Education Fresh F&V Consumption 9.690 0.600 0.658 < 0.001 White High Education Fast Food Consumption -5.878 0.363 0.342 < 0.001 White High Education Soda Consumption -4.346 0.036 0.035 < 0.001 White High Education BMI -4.100 28.955 27.768 < 0.001 White High Education % Overweight or Obese -11.105 0.717 0.638 < 0.001 White Low Fast Food Fresh F&V Consumption 6.028 0.625 0.663 < 0.001 White Low Fast Food Fast Food Consumption -6.611 0.359 0.335 < 0.001 White Low Fast Food Soda Consumption -16.363 0.039 0.033 < 0.001 White Low Fast Food BMI -0.644 28.147 27.966 < 0.001 White Low Fast Food % Overweight or Obese -2.373 0.662 0.647 < 0.001 White High Grocery Fresh F&V Consumption 1.655 0.634 0.644 < 0.001 White High Grocery Fast Food Consumption -5.004 0.368 0.349 < 0.001 White High Grocery Soda Consumption -5.319 0.038 0.036 < 0.001 White High Grocery BMI -0.104 28.239 28.209 0.082 White High Grocery % Overweight or Obese -1.373 0.665 0.656 < 0.001

11 Table e5: Effect sizes of all top/bottom quartiles matching experiments (Fig. e1).

Treatment Outcome % Change Trt. Mean Ctrl. Mean P-value (bootstrapping.)

High Income Fresh F&V Consumption 7.383 0.636 0.683 < 0.001 High Income Fast Food Consumption -9.366 0.345 0.312 < 0.001 High Income Soda Consumption -7.064 0.032 0.029 < 0.001 High Income BMI 0.689 27.175 27.362 < 0.001 High Income % Overweight or Obese 2.904 0.596 0.613 < 0.001 High Education Fresh F&V Consumption 13.977 0.582 0.664 < 0.001 High Education Fast Food Consumption -10.323 0.382 0.342 < 0.001 High Education Soda Consumption -12.293 0.039 0.034 < 0.001 High Education BMI -6.641 29.510 27.550 < 0.001 High Education % Overweight or Obese -15.904 0.739 0.621 < 0.001 High Grocery Fresh F&V Consumption 5.664 0.614 0.649 < 0.001 High Grocery Fast Food Consumption -11.619 0.379 0.335 < 0.001 High Grocery Soda Consumption -14.903 0.038 0.032 < 0.001 High Grocery BMI -0.646 28.564 28.379 < 0.001 High Grocery % Overweight or Obese -2.613 0.683 0.665 < 0.001 Low Fast Food Fresh F&V Consumption 8.893 0.613 0.668 < 0.001 Low Fast Food Fast Food Consumption -11.542 0.371 0.328 < 0.001 Low Fast Food Soda Consumption -30.282 0.042 0.029 < 0.001 Low Fast Food BMI -0.153 28.161 28.118 0.098 Low Fast Food % Overweight or Obese -0.384 0.656 0.654 0.140

12 a. b.

Fresh Fruits and Vegetables Consumption Fast Food Consumption

30 30

25 25

20 20

15 15

10 10

5 5

0 0

5 5 Relative Increase in Consumption (in %) Relative Decrease in Consumption (in %)

High Income High Grocery High Income High Grocery High Education Low Fast Food High Education Low Fast Food

c. d.

Soda Consumption % Overweight or Obese

30 30

25 25

20 20

15 15

10 10

5 5

0 0

5 5 Relative Decrease in Consumption (in %) Relative Decrease in Fraction Overweight or Obese (in %) High Income High Grocery High Income High Grocery High Education Low Fast Food High Education Low Fast Food

Figure e1: Matching experiments using quartile instead of median split. Note the consistent but increased effect sizes compared to Figure 4.

13 a. b.

Fresh Fruits and Vegetables Consumption Fast Food Consumption

17.5 17.5

15.0 15.0

12.5 12.5

10.0 10.0

7.5 7.5

5.0 5.0

2.5 2.5

0.0 0.0 Relative Increase in Consumption (in %) Relative Decrease in Consumption (in %)

2.5 2.5

High Income High Grocery High Income High Grocery High EducationLow Fast Food High EducationLow Fast Food

Low Electronics Stores Low Electronics Stores Low Countertop Installers Low Countertop Installers c. Low Waterproofing Services d. Low Waterproofing Services

Soda Consumption % Overweight or Obese

17.5 17.5

15.0 15.0

12.5 12.5

10.0 10.0

7.5 7.5

5.0 5.0

2.5 2.5

0.0 0.0 Relative Decrease in Consumption (in %)

2.5 2.5 Relative Decrease in Fraction Overweight or Obese (in %)

High Income High Grocery High Income High Grocery High EducationLow Fast Food High EducationLow Fast Food

Low Electronics Stores Low Electronics Stores Low Countertop Installers Low Countertop Installers Low Waterproofing Services Low Waterproofing Services Figure e2: Demonstration of discriminant validity of statistical approach. We measured the effect of null-treatments that should not have any impact on food consumption. We chose examples of null-treatments by selecting variables that had little correlation with study independent variables (income, education, grocery access, fast food access) and were plausibly unrelated to food consumption. This selection process lead to use of the fraction of countertop installers, electronics stores, and waterprooﬁng services nearby as measured through Yelp. Applying our analysis pipeline to these null-treatments, we found that all of these effect estimates were close to zero. This demonstrated that our statistical analysis approach did not produce measurements that it was not supposed to measure; that is, discriminant validity. 14 Table e6: USA fast food restaurants table used to classify participant food entries as fast food102. A list of popular pizza chains from the USA was appended to the list103.

A&W Restaurants Cinnabon Red Burrito Rogers Restaurants Chuck E. Cheese’s Murphy’s Pat’s Pizza Arby’s Claim Jumper The Habit Runza Saladworks CiCi’s Pizza Cottage Patxi’s Chicago Peter Arctic Circle Coco’s Halal Guys Sbarro Schlotzsky’s Inn Dion’s Discovery Piper Pie Five Arthurs Cold Stone Creamery Hardee’s Seattle’s Best Shake Zone Domino’s Pietro’s Pizza Pizza Atlanta Bread Cookout Huddle House Shack Skyline Chili Donatos Corner Pizza Factory Au Bon Pain Copeland’s In-N-Out Burger Sneaky Pete’s Sonic DoubleDave’s East of Pizza Fusion Pizza Auntie Anne’s Old Country Jack in the Box Spangles Steak Chicago Eatza Hut Pizza King Pizza Baja Fresh Culver’s Jack’s Family Escape Steak ’n Extreme Pizza Inn Pizza My Heart Bakers Square Dairy Queen Restaura Shake Stir Crazy Sub Fazoli’s Fellini’s Pizza Patron´ Pizza Blimpies el Tacos ts Jersey Mike’s Subs Station II Subway Fox’s Frank Pepe Ranch Pizza Bojangles DiBella’s Jimmy John’s Jim’s Swensen’s Swensons Gatti’s Gino’s Schmizza The Pizza Boston Market Dixie Chili and Deli Restaurants Johnny Taco Bell Taco Giordano’s Studio Pizzeria Venti Braum’s Don Pablo’s Rockets KFC Bueno Taco Cabana Godfather’s Regina Pizzeria Burger Chef Druther’s Kewpee Krispy Taco John’s Taco Grimaldi’s Grotto Rocky Rococo Burger King Dunkin’ Donuts Kreme L&L Mayo Taco Tico Taco Pizza Happy Joe’s Rosati’s Round Table Burger Street Eat’n Park Hawaiian Barbecue Time Twin Peaks Happy’s Hideaway Pizza Russo’s New Burgerville Eegee’s Lee Roy Selmon’s Umami Burger Home Run Inn York Pizze Captain D’s El Chico Lee’s Famous Lion’s Wendy’s Wetzel’s Hungry Howie’s ia Sal’s Pizza Carino’s Italian Grill El Pollo Loco Choice Long John Pretzels Whataburger Hunt Brothers Imo’s Sammy’s Carl’s Jr. El Taco Tote Silver’s Luby’s White Castle Pizza Jerry’s Jet’s Sarpino’s Carrows Elephant Bar McDonald’s Milo’s Wienerschnitzel John’s LaRosa’s Sbarro Charley’s Grilled Elevation Burger Moe’s Mooyah Mr. Wimpy Zaxby’s Ledo Little Caesars Shakey’s Subs Famous Dave’s Hero Mrs. Fields Zero’s Subs Zippy’s Lou Malnati’s Showbiz Checkers Farmer Boys Mrs. Winner’s America’s Incredible Marco’s Marion’s Sir Pizza Cheeburger Fatburger Chicken Arni’s Aurelio’s Mark’s Mazzio’s Snappy Tomato Cheeburger Firehouse Subs Biscuits Naugles Azzip Bearno’s Mellow Mushroom Straw Hat Chevys Five Guys Panera Bread Panda Bertucci’s Big MOD Pizza Toppers Chicken Express Freddy’s Express Penn Station Mama’s & Papa’s Monical’s Mountain Uncle Maddio’s Chick-ﬁl-A Freddies Pita Pit Popeyes Port Blackjack Blaze Mike’s Mr. Jim’s Unos Chronic Tacos Golden Chick of Subs Potbell Buddy’s Bullwinkle’s Noble Roman’s Old Upper Crust Pizzeria Chuck-A-Rama Good Times Quizno’s Raising California Pizza Chicago Pacpizza Valentino’s Church’s Great Steak Cane’s Rax Roast Casey’s General Pagliacci Papa Gino’s Vocelli Pizza Texas Chicken Green Burrito Beef Robeks Roy Stores Cassano’s Papa John’s Papa Your Pie

15 Table e7: USA soda list used to classify participant food entries as sugary sodas. The list was constructed using a list of America’s best-selling brands of Soda104, in addition to the generic terms such “Root Beer” and “Coca”.

Pepsi Dr Pepper Root Beer Coca Cola Sprite Coke Mountain Dew Fanta Coca

16 Table e8: Summary of High Income (MedianFamilyIncome > Median) matching experiment

(a) Sample sizes

Control Treated All 4911 4911 Matched 1358 4911 Unmatched 3553 0 Discarded 0 0

(b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.75 0.75 0.26 0.00 0.01 Education (% Without College Degree) 0.56 0.57 0.15 -0.01 -0.06 Grocery Distance (USDA lapophalfshare) 0.74 0.74 0.21 0.00 0.00 Yelp Fast Food % 0.06 0.06 0.06 0.00 -0.01 Median Family Income 97244.90 58991.91 8680.39 38253.00 4.41

17 Table e9: Summary of High Grocery (Grocery Access > Median) matching experiment

(a) Sample sizes

Control Treated All 4911 4911 Matched 2219 4911 Unmatched 2692 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.59 0.58 0.19 0.00 0.01 Median Family Income 76817.91 77021.36 30133.94 -203.45 -0.01 Education (% Without College Degree) 0.64 0.64 0.18 0.00 -0.01 Yelp Fast Food % 0.06 0.06 0.06 0.00 0.00 Grocery Distance (USDA lapophalfshare) 0.57 0.88 0.06 -0.31 -5.62

18 Table e10: Summary of High Education (% College Degrees > Median) matching experiment

(a) Sample sizes

Control Treated All 4911 4911 Matched 1481 4911 Unmatched 3430 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.73 0.73 0.27 0.00 0.01 Median Family Income 93203.36 88493.86 20530.36 4709.51 0.23 Grocery Distance (USDA lapophalfshare) 0.71 0.72 0.22 -0.01 -0.05 Yelp Fast Food % 0.06 0.06 0.06 0.00 0.02 Education (% Without College Degree) 0.53 0.75 0.05 -0.22 -4.79

19 Table e11: Summary of Low Fast Food (% Yelp Fast Food < Median) matching experiment

(a) Sample sizes

Control Treated All 4911 4911 Matched 1910 4911 Unmatched 3001 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.65 0.65 0.24 0.00 0.01 Median Family Income 86942.10 86597.30 31088.00 344.80 0.01 Education (% Without College Degree) 0.60 0.61 0.18 0.00 -0.02 Grocery Distance (USDA lapophalfshare) 0.66 0.67 0.23 -0.01 -0.05 Yelp Fast Food % 0.03 0.11 0.06 -0.08 -1.35

20 Table e12: Summary of High Income (MedianFamilyIncome > 75th Percentile) matching experiment

(a) Sample sizes

Control Treated All 2319 2321 Matched 215 2321 Unmatched 2104 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.91 0.91 0.18 0.00 0.01 Education (% Without College Degree) 0.47 0.48 0.13 -0.02 -0.11 Grocery Distance (USDA lapophalfshare) 0.71 0.70 0.22 0.01 0.04 Yelp Fast Food % 0.04 0.04 0.04 0.00 -0.06 Median Family Income 115443.80 48357.34 6021.64 67086.46 11.14

21 Table e13: Summary of High Grocery (Grocery Access > 75th Percentile) matching experiment

(a) Sample sizes

Control Treated All 2321 2319 Matched 743 2319 Unmatched 1578 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.69 0.68 0.23 0.01 0.03 Median Family Income 77354.53 77809.00 31756.28 -454.47 -0.01 Education (% Without College Degree) 0.62 0.62 0.19 0.00 -0.01 Yelp Fast Food % 0.04 0.04 0.04 0.00 -0.01 Grocery Distance (USDA lapophalfshare) 0.41 0.94 0.04 -0.52 -13.78

22 Table e14: Summary of High Education (% College Degrees > 75th Percentile) matching experiment. Note: Treatment samples dropped due to 0.3 STD caliper used to ensure 0.25 SMD balancing constraint.

(a) Sample sizes

Control Treated All 2321 2319 Matched 252 834 Unmatched 2069 1485 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.73 0.71 0.30 0.02 0.06 Median Family Income 81374.28 79453.49 12961.32 1920.79 0.15 Grocery Distance (USDA lapophalfshare) 0.70 0.70 0.22 0.00 -0.01 Yelp Fast Food % 0.06 0.06 0.06 0.00 0.02 Education (% Without College Degree) 0.48 0.83 0.04 -0.34 -9.25

23 Table e15: Summary of Low Fast Food (% Yelp Fast Food < 25th Percentile) matching experiment. Note: Treatment samples dropped due to 1.5 STD caliper used to ensure 0.25 SMD balancing constraint.

(a) Sample sizes

Control Treated All 2322 2319 Matched 504 2166 Unmatched 1818 153 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.78 0.77 0.27 0.01 0.04 Median Family Income 88631.94 83632.13 20794.81 4999.81 0.24 Education (% Without College Degree) 0.59 0.59 0.15 0.00 -0.02 Grocery Distance (USDA lapophalfshare) 0.63 0.66 0.23 -0.03 -0.14 Yelp Fast Food % 0.02 0.18 0.05 -0.16 -3.23

24 Table e16: Summary of Low Countertop Installation Services (% Yelp Countertop Installers < Median) matching experiment

(a) Sample sizes

Control Treated All 4913 4909 Matched 2829 4909 Unmatched 2084 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.50 0.50 0.05 0.00 0.01 Median Family Income 77255.32 77443.29 28361.82 -187.96 -0.01 Education (% Without College Degree) 0.65 0.66 0.18 0.00 0.00 Grocery Distance (USDA lapophalfshare) 0.72 0.72 0.24 0.00 0.00 Yelp Fast Food % 0.09 0.09 0.08 0.00 0.01 Yelp Countertop Installers % 0.00 0.00 0.00 0.00 -1.61

25 Table e17: Summary of Low Electronics Stores (% Yelp Electronics Stores < Median) matching experiment

(a) Sample sizes

Control Treated All 4912 4910 Matched 2741 4910 Unmatched 2171 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.52 0.52 0.09 0.00 0.00 Median Family Income 74567.18 74306.84 25820.76 260.34 0.01 Education (% Without College Degree) 0.68 0.68 0.16 0.00 0.00 Grocery Distance (USDA lapophalfshare) 0.71 0.71 0.22 0.00 -0.01 Yelp Fast Food % 0.08 0.08 0.07 0.00 0.00 Yelp Electronics Stores % 0.00 0.00 0.00 0.00 -1.59

26 Table e18: Summary of Low Waterprooﬁng Services (% Yelp Waterprooﬁng Services < Median) matching experiment

(a) Sample sizes

Control Treated All 4913 4909 Matched 2876 4909 Unmatched 2037 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.50 0.50 0.02 0.00 0.00 Median Family Income 76371.54 76438.89 28071.29 -67.35 0.00 Education (% Without College Degree) 0.66 0.66 0.17 0.00 0.00 Grocery Distance (USDA lapophalfshare) 0.74 0.74 0.21 0.00 -0.01 Yelp Fast Food % 0.09 0.09 0.08 0.00 0.01 Yelp Waterprooﬁng Services % 0.00 0.00 0.00 0.00 -1.48

27 Table e19: Summary of Black-majority Zip Code High Income (MedianFamilyIncome > Median) matching experiment

(a) Sample sizes

Control Treated All 317 42 Matched 30 42 Unmatched 287 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.36 0.35 0.25 0.02 0.07 Education (% Without College Degree) 0.65 0.65 0.09 -0.01 -0.08 Grocery Distance (USDA lapophalfshare) 0.68 0.67 0.22 0.01 0.02 Yelp Fast Food % 0.04 0.03 0.03 0.00 0.09 Median Family Income 88944.14 59779.41 8039.94 29164.73 3.63

28 Table e20: Summary of Hispanic-majority Zip Code High Income (MedianFamilyIncome > Me- dian) matching experiment

(a) Sample sizes

Control Treated All 482 67 Matched 51 67 Unmatched 431 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.35 0.31 0.25 0.04 0.16 Education (% Without College Degree) 0.70 0.72 0.12 -0.02 -0.16 Grocery Distance (USDA lapophalfshare) 0.62 0.63 0.20 -0.01 -0.05 Yelp Fast Food % 0.07 0.07 0.05 0.00 0.00 Median Family Income 82812.73 56050.33 8534.43 26762.40 3.14

29 Table e21: Summary of White-majority Zip Code High Income (MedianFamilyIncome > Median) matching experiment

(a) Sample sizes

Control Treated All 3421 4277 Matched 1023 4277 Unmatched 2398 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.78 0.78 0.25 0.00 0.01 Education (% Without College Degree) 0.55 0.56 0.15 -0.01 -0.07 Grocery Distance (USDA lapophalfshare) 0.76 0.76 0.19 0.00 -0.01 Yelp Fast Food % 0.06 0.06 0.06 0.00 -0.01 Median Family Income 98014.79 59878.82 8719.97 38135.97 4.37

30 Table e22: Summary of Black-majority Zip Code High Grocery (Grocery Access > Median) matching experiment

(a) Sample sizes

Control Treated All 100 259 Matched 65 259 Unmatched 35 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.78 0.77 0.15 0.01 0.08 Median Family Income 51410.82 51925.49 14357.87 -514.66 -0.04 Education (% Without College Degree) 0.76 0.77 0.08 0.00 -0.03 Yelp Fast Food % 0.05 0.05 0.05 0.00 -0.08 Grocery Distance (USDA lapophalfshare) 0.53 0.86 0.04 -0.33 -9.43

31 Table e23: Summary of Hispanic-majority Zip Code High Grocery (Grocery Access > Median) matching experiment

(a) Sample sizes

Control Treated All 78 471 Matched 66 471 Unmatched 12 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.87 0.87 0.09 0.00 0.01 Median Family Income 52218.86 53068.49 13918.72 -849.63 -0.06 Education (% Without College Degree) 0.82 0.83 0.08 -0.01 -0.12 Yelp Fast Food % 0.06 0.06 0.05 0.00 0.01 Grocery Distance (USDA lapophalfshare) 0.47 0.88 0.05 -0.41 -7.86

32 Table e24: Summary of White-majority Zip Code High Grocery (Grocery Access > Median) matching experiment

(a) Sample sizes

Control Treated All 4494 3204 Matched 1741 3204 Unmatched 2753 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.51 0.50 0.19 0.00 0.02 Median Family Income 84030.48 84275.98 31002.90 -245.50 -0.01 Education (% Without College Degree) 0.59 0.59 0.18 0.00 -0.01 Yelp Fast Food % 0.07 0.07 0.07 0.00 0.00 Grocery Distance (USDA lapophalfshare) 0.62 0.89 0.05 -0.27 -4.92

33 Table e25: Summary of Black-majority Zip Code High Education (% College Degrees > Median) matching experiment

(a) Sample sizes

Control Treated All 285 74 Matched 48 74 Unmatched 237 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.46 0.45 0.29 0.01 0.02 Median Family Income 70932.69 71048.45 22123.91 -115.77 -0.01 Grocery Distance (USDA lapophalfshare) 0.59 0.60 0.28 -0.01 -0.04 Yelp Fast Food % 0.04 0.04 0.04 0.00 0.00 Education (% Without College Degree) 0.62 0.77 0.04 -0.15 -3.39

34 Table e26: Summary of Hispanic-majority Zip Code High Education (% College Degrees > Me- dian) matching experiment

(a) Sample sizes

Control Treated All 488 61 Matched 43 61 Unmatched 445 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.33 0.28 0.22 0.05 0.21 Median Family Income 71569.13 67609.51 16348.80 3959.62 0.24 Grocery Distance (USDA lapophalfshare) 0.51 0.52 0.27 -0.01 -0.04 Yelp Fast Food % 0.05 0.05 0.04 0.00 -0.06 Education (% Without College Degree) 0.61 0.80 0.05 -0.19 -3.79

35 Table e27: Summary of White-majority Zip Code High Education (% College Degrees > Median) matching experiment. Note: Treatment samples dropped due to 2.1 STD caliper used to ensure 0.25 SMD balancing constraint.

(a) Sample sizes

Control Treated All 3491 4207 Matched 1114 4102 Unmatched 2377 105 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.75 0.75 0.26 0.00 0.01 Median Family Income 92654.95 88357.13 17926.24 4297.82 0.24 Grocery Distance (USDA lapophalfshare) 0.74 0.76 0.17 -0.01 -0.07 Yelp Fast Food % 0.07 0.07 0.06 0.00 0.00 Education (% Without College Degree) 0.53 0.75 0.05 -0.22 -4.59

36 Table e28: Summary of Black-majority Zip Code Low Fast Food (% Yelp Fast Food < Median) matching experiment

(a) Sample sizes

Control Treated All 100 259 Matched 70 259 Unmatched 30 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.77 0.76 0.16 0.00 0.03 Median Family Income 54812.73 54366.39 18020.16 446.34 0.02 Education (% Without College Degree) 0.75 0.74 0.10 0.01 0.09 Grocery Distance (USDA lapophalfshare) 0.57 0.60 0.22 -0.02 -0.11 Yelp Fast Food % 0.03 0.11 0.07 -0.08 -1.21

37 Table e29: Summary of Hispanic-majority Zip Code Low Fast Food (% Yelp Fast Food < Median) matching experiment

(a) Sample sizes

Control Treated All 252 297 Matched 135 297 Unmatched 117 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.61 0.61 0.19 0.00 0.02 Median Family Income 52130.57 51853.43 13411.40 277.14 0.02 Education (% Without College Degree) 0.81 0.82 0.09 -0.01 -0.11 Grocery Distance (USDA lapophalfshare) 0.45 0.45 0.26 0.00 -0.01 Yelp Fast Food % 0.03 0.10 0.06 -0.07 -1.20

38 Table e30: Summary of White-majority Zip Code Low Fast Food (% Yelp Fast Food < Median) matching experiment

(a) Sample sizes

Control Treated All 4188 3510 Matched 1362 3510 Unmatched 2826 0 Discarded 0 0 (b) Summary of balance for matched data

Means Treated Means Control SD Control Mean Diff Std. Mean Difference distance 0.65 0.65 0.26 0.00 0.01 Median Family Income 95359.35 94854.08 29365.91 505.27 0.02 Education (% Without College Degree) 0.56 0.57 0.17 0.00 -0.01 Grocery Distance (USDA lapophalfshare) 0.71 0.72 0.20 -0.01 -0.05 Yelp Fast Food % 0.03 0.11 0.06 -0.08 -1.37